Unverified50% confidenceFactExact time
BDH (Dragon Hatchling) combines linear attention in high-dimensional neural space with a low-rank GPU implementation, where its recurrent attention state is an N×D matrix (N much greater than D) rather than an N×N connectivity matrix.
1
Sources
50%
Confidence
Long-term
Relevance
10/6/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLongformer、BigBird通过局部滑动窗口注意力加少量全局token的混合注意力策略，将计算复杂度从O(n²)降至O(n)67% similarUnverifiedMobile NPUs are better at handling fixed-shape CNN inference, while the self-attention mechanism of LLMs has a dynamically shaped computational graph, which means CPU+GPU hybrid inference is often more mature than pure NPU solutions.66% similarUnverifiedIn the standard attention mechanism, the Query matrix and Key matrix perform matrix multiplication to obtain an n×n attention score matrix, with a computational complexity of O(n²·d).66% similarUnverifiedLinformer 将注意力矩阵低秩近似为 O(n) 复杂度，FlashAttention 通过重排计算顺序降低 GPU 内存带宽占用，Longformer 引入滑动窗口注意力65% similarUnverifiedLinear attention is better suited to NPU chips due to the static graph characteristics of its fixed computation, whereas traditional attention uses dynamic graphs that require chips to frequently switch scheduling.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/981676API
curl https://kongchang.com/api/v1/knowledge/claims/981676MCP
get_claim(id=981676)